Deep learning on the fusion of chemical sequences and molecular grids for ligand-based virtual screening

Computational techniques have been widely applied in modern drug discovery to reduce cost and time. As a crucial component of computational drug discovery, predicting compound-protein interactions is becoming increasingly prevalent. Virtual screening serves as a cost-effective tool for predicting such interactions. However, current models often require input from both compounds and proteins, which is unfriendly for scenarios where the protein information is not valid. In this work, we propose a ligand-based prediction approach that leverages only the SMILES sequences and 3D grids of compounds in target-specific tasks. By learning these features through a cross-attention mechanism, our model can capture high-level structural features of the compounds in the prediction tasks. Experimental validation on the DUD-E dataset demonstrates that our model achieves competitive performance in both accuracy and efficiency. Particularly, it performs decently when large proteins are involved.